<rss xmlns:atom="http://www.w3.org/2005/Atom" version="2.0">
    <channel>
        <title>Deepresearch - 标签 - vllbc02&#39;s blogs</title>
        <link>https://blog.vllbc.top/tags/deepresearch/</link>
        <description>Deepresearch - 标签 - vllbc02&#39;s blogs</description>
        <generator>Hugo -- gohugo.io</generator><language>zh-CN</language><copyright>This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License.</copyright><lastBuildDate>Tue, 23 Sep 2025 00:00:00 &#43;0000</lastBuildDate><atom:link href="https://blog.vllbc.top/tags/deepresearch/" rel="self" type="application/rss+xml" /><item>
    <title>DeepDive：Advancing Deep Search Agents with Knowledge Graphs and Multi-Turn RL</title>
    <link>https://blog.vllbc.top/deepdiveadvancing-deep-search-agents-with-knowledge-graphs-and-multi-turn-rl/</link>
    <pubDate>Tue, 23 Sep 2025 00:00:00 &#43;0000</pubDate>
    <author>vllbc</author>
    <guid>https://blog.vllbc.top/deepdiveadvancing-deep-search-agents-with-knowledge-graphs-and-multi-turn-rl/</guid>
    <description><![CDATA[<p>deepresearch的工作，解决的问题是ds训练中数据短缺的问题，提出的创新主要是数据合成，方法类似于逆向工程和<a
href="WebExplorer：Explore%20and%20Evolve%20for%20Training%20Long-Horizon%20Web%20Agents.md">WebExplorer：Explore
and Evolve for Training Long-Horizon Web
Agents</a>，本文的方法是在知识图谱上面随机探索，形成复杂推理路径，并且对实体进行模糊化，最后将路径作为问题，路径终点作为答案，并且还过滤难度，过于简单的直接抛弃。</p>]]></description>
</item>
<item>
    <title>SFR-DeepResearch：Towards Effective Reinforcement Learning for Autonomously Reasoning Single Agents</title>
    <link>https://blog.vllbc.top/sfr-deepresearchtowards-effective-reinforcement-learning-for-autonomously-reasoning-single-agents/</link>
    <pubDate>Wed, 10 Sep 2025 00:00:00 &#43;0000</pubDate>
    <author>vllbc</author>
    <guid>https://blog.vllbc.top/sfr-deepresearchtowards-effective-reinforcement-learning-for-autonomously-reasoning-single-agents/</guid>
    <description><![CDATA[<p>
将多轮对话用单轮训练，类似verl-agent的做法，当上下文较长时会总结成summary。</p>
<hr />
<h3
id="论文深度解读从思考模型到行动专家的进化之路">论文深度解读：从“思考模型”到“行动专家”的进化之路</h3>
<p>这篇来自 Salesforce AI Research
的论文，核心目标是解决一个当前大模型（LLM）智能体领域极具挑战性的问题：如何高效地训练一个<strong>自主单智能体（Autonomous
Single-Agent）</strong>，使其能够在没有人类持续干预的情况下，独立完成需要广泛搜索、深度整合信息和复杂推理的<strong>深度研究（Deep
Research, DR）</strong>任务。</p>]]></description>
</item>
<item>
    <title>ReSearch：Learning to Reason with Search for LLMs via Reinforcement Learning</title>
    <link>https://blog.vllbc.top/researchlearning-to-reason-with-search-for-llms-via-reinforcement-learning/</link>
    <pubDate>Tue, 09 Sep 2025 00:00:00 &#43;0000</pubDate>
    <author>vllbc</author>
    <guid>https://blog.vllbc.top/researchlearning-to-reason-with-search-for-llms-via-reinforcement-learning/</guid>
    <description><![CDATA[<p>和<a
href="Search-R1：Training%20LLMs%20to%20Reason%20and%20Leverage%20Search%20Engines%20with%20Reinforcement%20Learning.md">Search-R1：Training
LLMs to Reason and Leverage Search Engines with Reinforcement
Learning</a>的区别就是多了个格式奖励。</p>
<hr />
<figure>

<figcaption aria-hidden="true">image.png</figcaption>
</figure>
<p>这篇论文的核心贡献在于提出了一种名为 <strong>ReSearch</strong>
的新颖框架，它<strong>利用强化学习（Reinforcement
Learning）来训练大语言模型（LLM）在推理过程中主动、智能地使用搜索工具</strong>，而整个训练过程完全不需要人工标注的推理步骤数据。这标志着从传统的、相对固化的“检索-增强-生成”（RAG）模式，向一种更动态、更智能的“边思考边搜索”的代理（Agent）行为模式的重大转变。</p>]]></description>
</item>
<item>
    <title>AI-SEARCHPLANNER：MODULAR AGENTIC SEARCH VIA PARETO-OPTIMAL MULTI-OBJECTIVE REINFORCEMENT LEARNING</title>
    <link>https://blog.vllbc.top/ai-searchplannermodular-agentic-search-via-pareto-optimal-multi-objective-reinforcement-learning/</link>
    <pubDate>Sat, 30 Aug 2025 00:00:00 &#43;0000</pubDate>
    <author>vllbc</author>
    <guid>https://blog.vllbc.top/ai-searchplannermodular-agentic-search-via-pareto-optimal-multi-objective-reinforcement-learning/</guid>
    <description><![CDATA[<p>将与搜索工具交互与答案整理分离，也就是术业有专攻。根据最终答案的正确奖励来优化搜索llm。</p>
<hr />
<h3 id="论文深度解读">论文深度解读</h3>
<p>首先，让我们对这篇论文进行一个全面的剖析。当前，利用大语言模型（LLM）与搜索引擎结合来解决复杂问题已成为主流。一种常见的方法是使用<strong>强化学习（RL）</strong>来训练一个模型，让它学会如何与搜索引擎进行多轮交互，这个过程我们称之为<strong>智能体搜索（Agentic
Search）</strong>。然而，论文敏锐地指出了现有方法的症结所在：</p>]]></description>
</item>
<item>
    <title>SSRL：SELF-SEARCH REINFORCEMENT LEARNING</title>
    <link>https://blog.vllbc.top/ssrlself-search-reinforcement-learning/</link>
    <pubDate>Mon, 18 Aug 2025 00:00:00 &#43;0000</pubDate>
    <author>vllbc</author>
    <guid>https://blog.vllbc.top/ssrlself-search-reinforcement-learning/</guid>
    <description><![CDATA[<p>相较于<a
href="Search-R1：Training%20LLMs%20to%20Reason%20and%20Leverage%20Search%20Engines%20with%20Reinforcement%20Learning.md">Search-R1：Training
LLMs to Reason and Leverage Search Engines with Reinforcement
Learning</a>的使用真实的搜索引擎和<a
href="ZEROSEARCH：Incentivize%20the%20Search%20Capability%20of%20LLMs%20without%20Searching.md">ZEROSEARCH：Incentivize
the Search Capability of LLMs without
Searching</a>使用虚拟的训练的大模型作为搜索引擎，这篇论文直接让自己当搜索引擎，更节省成本。</p>]]></description>
</item>
<item>
    <title>ParallelSearch：Train your LLMs to Decompose Query and Search Sub-queries in Parallel with Reinforcement Learning</title>
    <link>https://blog.vllbc.top/parallelsearchtrain-your-llms-to-decompose-query-and-search-sub-queries-in-parallel-with-reinforcement-learning/</link>
    <pubDate>Fri, 15 Aug 2025 00:00:00 &#43;0000</pubDate>
    <author>vllbc</author>
    <guid>https://blog.vllbc.top/parallelsearchtrain-your-llms-to-decompose-query-and-search-sub-queries-in-parallel-with-reinforcement-learning/</guid>
    <description><![CDATA[<p>和<a
href="TURA：Tool-Augmented%20Unified%20Retrieval%20Agent%20for%20AI%20Search.md">TURA：Tool-Augmented
Unified Retrieval Agent for AI
Search</a>一样，考虑的是并行的工具调用，不过TURA是整个系统，通过DAG来编排工具调用（外部LLM）。而本文研究的是提高大模型本身并行调用工具的能力。从而节省很多时间。</p>]]></description>
</item>
<item>
    <title>TURA：Tool-Augmented Unified Retrieval Agent for AI Search</title>
    <link>https://blog.vllbc.top/turatool-augmented-unified-retrieval-agent-for-ai-search/</link>
    <pubDate>Fri, 15 Aug 2025 00:00:00 &#43;0000</pubDate>
    <author>vllbc</author>
    <guid>https://blog.vllbc.top/turatool-augmented-unified-retrieval-agent-for-ai-search/</guid>
    <description><![CDATA[<p>问题分解-&gt;工具DAG编排（规划）-&gt;蒸馏小型执行器。</p>
<hr />
<h3 id="论文综合解读">论文综合解读</h3>
<p>这篇发表于2025年的论文，由百度公司的研究团队撰写，提出了一种名为
<strong>TURA (Tool-Augmented Unified Retrieval Agent for AI
Search)</strong> 的新型AI搜索框架。TURA的核心目标是突破当前主流的、基于
<strong>检索增强生成 (Retrieval-Augmented Generation, RAG)</strong>
技术的AI搜索在处理真实世界复杂查询时的根本性限制。</p>]]></description>
</item>
</channel>
</rss>
